1
Diagnosis, not just pass/fail
A Deep Verification Report with per-item scores, cited evidence and risk flags.
Verification Criteria v1
These are the criteria GAV verifies against. They are published in full, before you apply.
WIA-GAV-001 through 005 are criteria numbers within the WIA Standards system (wiastandards.com). Every GAV assessment cites these published standard numbers, so a verification is made against a standard rather than against ad hoc criteria.
A. AI Processing System — WIA-GAV-001
A1. An AI processing system (an algorithm/program designed to learn or infer from data to produce predictions, classifications, or generations) is embedded in the product or service.
A2. The system demonstrates learning and/or inference characteristics (recognized open-source/commercial AI systems qualify; proprietary systems require evidence such as training-variance or out-of-sample inference demonstrations).
A3. Input data produces meaningful outputs (prediction, generation, classification) through the system.
B. AI Functionality — WIA-GAV-002
B1. The architecture diagram clearly locates the AI component, and it is integrated with the product's operation.
B2. The AI demonstrably contributes to functionality, convenience, accessibility, or efficiency (one of five recognized purposes: new capability, quality improvement, user convenience, accessibility, resource efficiency).
B3. AI-powered core functions operate normally across varied inputs.
C. External AI Integration — WIA-GAV-003 (where applicable)
C1. For external AI services, call logs or API credentials are verifiable.
D. SaaS & Combined Products — WIA-GAV-004 (specific to GAV)
D1. For SaaS: the AI processing is part of the service's own processing pipeline, not a mere pass-through of a third-party AI interface.
D2. For AI+hardware combined products: removing the AI component would disable or materially degrade the product's core function.
F. Inclusive AI & Transparency — WIA-GAV-005 (specific to GAV)
F is not a pass/fail gate. It is a mandatory section of the Deep Verification Report: every item receives a grade and improvement recommendations, annexed to the certificate. GAV does not only look at whether a product passes — it looks at whether the product can be used by everyone.
F1. Accessibility scan — an automated, WCAG-based accessibility report of the product's web/app surface, where such a surface exists.
F2. AI transparency notice — for generative and high-impact AI products, whether a user-facing notice or indication of AI use exists. For products in the Korean market, alignment with the methods set out in Article 31 of the AI Framework Act and Article 23 of its Enforcement Decree (product, contract, on-screen indication, machine-readable marking) is recorded in the report; for global products, alignment with EU AI Act transparency principles is recorded.
F3. Vulnerable-user consideration — consideration for users across the ten categories defined in Article 1-2 of the Enforcement Decree (persons with disabilities; persons aged 65 or over; basic livelihood and near-poverty recipients; jobseekers eligible for benefits; women with career experience returning to work; farmers and fishers; members of multicultural families; North Korean defectors; recipients of single-parent family support; and others so recognized). Assessed on multiple input methods, readability and simplicity of operation, and whether the product falls under the 'accessibility improvement' purpose type — based on the applicant's own description together with supporting evidence (screens, settings, manuals).
Evidence standard
GAV states its evidence requirements openly, so that you know what to prepare.
- Items under A through C are not accepted on written description alone. At least one form of verifiable evidence is required: logs, reproducible test results, a weight-file hash, or API call records.
- Inconsistency across the application, architecture diagram, specification and manual is in itself grounds for a supplementary request.
- Out-of-sample tests are run to scenarios that GAV generates and provides; the applicant submits the results. Scenarios are not self-selected.
Depth of Verification — Deeper by Design
GAV maps to the published domestic criteria, but the method of verification differs. These seven layers are what that means in practice.
2
Cross-document consistency
Full cross-checking between the application, architecture diagram, specification and manual.
3
Recognized-AI-system auto-matching
Module names in the specification are matched against recognized AI system lists; modules outside the list receive automatic guidance on evidence requirements.
4
Evidence-level testing
Out-of-sample test scenarios are generated and provided, and API call log structures are analysed. Track D includes a demonstration script and checklist for the on-site verification stage.
5
Accessibility scan
A WCAG-based accessibility report on the product's web/app surface.
6
International mapping
A reference table to international frameworks and EU AI Act provisions.
7
Actionable remediation
For each item that falls short, guidance on how to correct and resubmit.
Not verifiable — examples
These patterns do not meet the criteria, however they are labelled.
Generic hardware with a pre-installed third-party AI app.
An LLM wired only to help or FAQ content.
A thin pass-through of an external AI prompt window.
Hard-coded rules labelled as "AI" — IF-based recommendations, simple statistics, scripted chatbots, keyword matching.